FinOps skills

CloudQuell ships a set of FinOps skills — packaged workflow runbooks in the open Agent Skills format — that teach your AI agent how to run proven FinOps workflows on top of the CloudQuell MCP server. Install a skill once and your agent knows the whole routine: which CloudQuell tools to call with which parameters, what to double-check, when to stop and ask you, and how to present the result.

Skills are instructions, not access. They only work once the CloudQuell connector is set up, and they never bypass CloudQuell’s gates: tool calls require a paid plan, and any change additionally requires the admin:write scope plus per-call confirmation — the skills themselves instruct the agent to get your explicit approval before any write.

The skills

Skill	What your agent learns	Writes?
Spend review	Monthly/weekly cost review: month-over-month deltas, top movers, service/account/tag breakdowns, direct AI-provider spend	No
Anomaly triage	Detect unusual spend, investigate to the resource level, dismiss noise with an audit trail or escalate real spikes	Dismissals only
Budget guardrails	Budget health & forecast checks, coverage audit, creating/tuning budgets, wiring and testing notification channels	Yes
Savings actions	Prioritize savings recommendations by value/effort/confidence, track them to banked savings, plan RI/Savings Plan renewals	Status updates
Allocation hygiene	Tag-coverage and cost-center allocation audit with a concrete, ranked remediation plan	No

Each skill is a folder containing a SKILL.md runbook, validated against the live tool surface. The read-only skills give full value on a read-only connection; the writing skills degrade gracefully to analysis + proposals if admin:write wasn’t granted.

Download

• All five skills (bundle): cloudquell-finops-skills.zip
• Individual skills: spend-review · anomaly-triage · budget-guardrails · savings-actions · allocation-hygiene

Install

Claude Code

Unzip the bundle into your project’s .claude/skills/ (shared with your team via the repo) or ~/.claude/skills/ (personal, all projects):

curl -sL https://docs.cloudquell.com/downloads/skills/cloudquell-finops-skills.zip -o /tmp/cq-skills.zip
unzip /tmp/cq-skills.zip -d ~/.claude/skills/

Add the connector if you haven’t:

claude mcp add --transport http cloudquell https://mcp.cloudquell.com/mcp

Claude picks the right skill automatically when a request matches — “why did our AWS bill spike?” engages anomaly triage — or invoke one explicitly, e.g. /cloudquell-spend-review.

Claude.ai / Claude Desktop

1. Add the CloudQuell connector: Settings → Connectors → Add custom connector with URL https://mcp.cloudquell.com/mcp (details in Connecting).
2. Upload skills under Settings → Capabilities → Skills, one individual skill zip at a time (the uploader expects one skill per zip).
3. In a chat, ask a matching question — the skill engages on its own.

Other agents

Any agent that supports the Agent Skills format can use these: unzip and point the agent at the skill folders. The tool names referenced in each SKILL.md are the CloudQuell MCP tool names as served by the endpoint, so they work regardless of any client-side prefixing.

What a skill run looks like

Asking “are we on budget this month?” with the budget-guardrails skill installed: the agent calls get_budget_status, leads with anything over or forecast-breaching, audits coverage against your actual top spend via query_costs, and comes back with a status table plus proposed changes — each with exact parameters — waiting for your go-ahead before creating or modifying anything.

Good to know

• Skills complement — not replace — the example prompts & workflows: the examples show what’s possible in one-off questions; skills make the full routine repeatable and consistent.
• Skills are versioned with the MCP tool surface. If a tool gains parameters or a new tool ships, updated skill zips land here — reinstalling is just replacing the folder. See the Changelog.
• On the Free plan the connector works but every tool call returns upgrade_required — the skills detect this and explain rather than fail cryptically. See Access, limits & data scope.